There’s an order that actually works for AI training work, and it runs backwards from how most job hunting goes: get registered before you polish anything.
Register first, optimise second
Most of this work doesn’t come from browsing a list of projects and applying to the ones you like. Once you’ve registered on a platform and uploaded a CV and LinkedIn profile, the platform’s own matching system starts looking for a fit on your behalf. When it finds one, you’ll often hear about it as an emailed “instant offer” rather than by going and searching for it yourself.
So get your name into the system before you polish anything. Register on the reputable platforms that are relevant to your background and that you are comfortable using – for example Mercor, micro1, Turing or AfterQuery -mdash; so their matching engines have you in the pool as early as possible. Only after that should you go back and spend proper time on your CV, LinkedIn and platform profile, so the offers that come back are ones worth taking. Most of this is part-time, remote work that fits around an existing job, and the faster you’re registered, the sooner it starts finding you rather than the other way round.
Types of AI training projects
Once you’re registered, it’s worth understanding what you’ll actually be asked to do. “AI training work” isn’t one job. Mercor, micro1 and Turing route contractors into different tiers and specialisms depending on background and credentials, and the project you land determines the pay, the pace, and how much of your professional expertise gets used.
Seven types of AI training task
We looked at our own live roles database — every open contract we track across Mercor, micro1, Turing and AfterQuery, tagged by the work involved — rather than guess at this from the outside. Almost every role, whatever profession sits behind it, comes down to one of seven task types. They’re listed below, biggest category first, with the pay range currently being advertised for each.
Market snapshot: role counts and pay ranges below reflect ACJ’s tracked live-role data as checked on 5 September 2026. Listings, rates and category counts change frequently; use the linked live-role pages for the current position.
- Model evaluation and comparison. The biggest category by some distance. Judging or ranking AI outputs against each other or against a rubric: which of two answers is better, does a video look realistic, does a legal argument hold up. $10–250/hour depending on the domain — this is where professional judgement is worth the most. 140 live roles →
- Expert reasoning and QA. Not judging someone else’s answer but originating one: writing step-by-step solutions, designing novel problems, building the reasoning datasets models are trained on. Common in physics, maths, biology, finance and law. $10–250/hour. 2 live roles →
- Annotation and labelling. Tagging, classifying and validating structured data: extracting fields from documents, building knowledge graphs, flagging edge cases. $8–200/hour. The low end is generic labelling; the high end needs real domain literacy, tax forms or clinical notes rather than stock photos. 33 live roles →
- Transcription and audio QA. Verifying transcripts, diarisation (working out who said what) and audio fidelity, often for a specific language or accent. $6–95/hour. Fluency in a less common language pushes you toward the top. 6 live roles →
- Voice, video and physical data capture. Recording voice-overs, filming yourself narrating or performing a task, or demonstrating physical actions for robotics data (more on that below). $6–65/hour for most of it, though specialist voice acting and physical demonstration work can run higher. 35 live roles →
- Software and agent benchmarking. Building and running evaluation tasks for coding models and AI agents: designing benchmarks from real code changes, debugging model output, building data pipelines. $30–170/hour, aimed at working software and ML engineers. 91 live roles →
- Translation and localisation. The smallest category by volume, but often the best-paid per hour for the right language pair, because supply is genuinely scarce. $15–95/hour, with rarer languages — we’ve tracked live roles for Sherpa and Tibetan — at the top of that range. 1 live role →
The link after each one goes to that category’s live jobs page with a running count, so you can check the current numbers yourself rather than take our word for it.
Notice how wide the pay range is within a single task. $10–250/hour for evaluation work isn’t a typo — it’s the gap between generalist labelling and a licensed specialist judging outputs in their own field. And most of these tasks aren’t tied to one profession: model evaluation and expert reasoning show up across finance, law, medicine and physics alike. The task matters at least as much as the job title on the listing.
Pay tiers: generalist, mid-level and expert
Underneath the task types sits a second layer: platforms sort work into three pay bands, and knowing which one you’re being matched to changes what to expect from the job.
- Generalist / entry-level — data labelling, image tagging, work that needs no specialist background. $12–25/hour.
- Mid-level — demonstrated fluency in a domain (writing, a subject area, a technical skill) without formal credentials. $25–53/hour.
- Expert / specialist — verifiable professional credentials or 5–10+ years of domain experience: licensed clinicians, lawyers, senior engineers, PhDs and post-docs. $75–200+/hour. Supply is genuinely limited at this level, and the rate reflects it.
For NHS and other healthcare professionals, this is the tier most readers here should be aiming for. Clinical registration and years of practice are exactly the “verifiable professional credentials” that route you into expert-tier projects rather than a generalist labelling queue.
Niche work: it isn’t all chatbot conversations
Most people picture AI training as rating chatbot answers. A fair amount of it is stranger, and more physical, than that. Robotics is the clearest example, and it isn’t a future prediction — it’s live in our tracked roles now. As checked on 5 September 2026, micro1 was advertising paid contracts where people record themselves doing ordinary household tasks while wearing a gripper-style device, VR headset and hand-tracking equipment, to train physical robots rather than language models. A second, lower-paid role captures the same kind of footage with nothing more than a compatible smartphone and a head-mounted camera. Neither needs any robotics background. What’s being bought is someone reliably performing an everyday physical task while wearing the capture equipment correctly.
Other contractors work the annotation side of the same pipeline: labelling that footage frame by frame with temporal segments, hand and object interactions, and step-by-step descriptions of what’s happening. It’s the annotation task type from above, just applied to video of a person rather than a document.
Robotics isn’t the only corner like this. Our tracked roles have also included presenting yourself on camera narrating a “walk and talk” tour of an interesting local space, building and curating knowledge graphs for a content-understanding platform, and paid rare-language work — live contracts specifically for Sherpa and Tibetan speakers, at rates well above the generalist floor, simply because so few people can do it.
If a platform’s project list looks irrelevant to your field, check back periodically rather than writing it off. New niches open up as models tackle new domains. “Expert” doesn’t only mean expert-in-computing — it can mean a tradesperson, a warehouse picker, a rare-language speaker, or someone who can demonstrate a physiotherapy exercise correctly on camera.
Permanent staff roles at the platforms themselves
Separate from all of the above: Mercor, micro1 and Turing are companies in their own right, and they hire permanent, salaried staff to build and run the platform — engineers, product managers, recruiters, researchers, operations people. These are ordinary full-time jobs, advertised on the company’s own careers page rather than inside the contractor project queue, and you apply for them the normal way: CV, interviews, an offer with a salary and equity.
Most of these roles are office-based, in hubs like San Francisco, New York or London, rather than remote, and pay is quoted as an annual salary rather than an hourly rate. If that’s genuinely what you’re after, rather than the flexible, part-time contractor work most of this site covers, go to the platform’s own careers page directly. It’s a completely separate application route from everything else described here.
Which tier should you aim for?
If you’re weighing up whether clinical AI training work is realistic for your background, the Am I suited? page goes into that in more detail. As a rule of thumb: registered professionals with several years of practice should apply to expert-tier projects directly rather than starting in a generalist queue. Platforms actively want your credentials, and generalist rates aren’t worth your time.
Frequently asked questions
Can I do both generalist and expert projects on the same platform?
Often, yes. Being matched to expert-tier work doesn’t lock you out of other projects, though most people who qualify for expert rates have little reason to pick up generalist tasks — the pay gap is too wide to make sense of it.
Do robotics and other niche projects pay differently to standard expert work?
Rates follow the same tier logic as any other project. What matters is the skill or credential being bought, not the subject matter. A physical demonstration role staffed by a qualified tradesperson can sit at expert rates; annotating the resulting footage without specialist knowledge typically sits closer to mid-level.
Should I apply to a platform’s careers page instead of its contractor projects?
Only if you specifically want a full-time, office-based, salaried role at the company itself. If you want flexible, remote, project-based AI training work around your existing job, apply as a contractor through the platform’s expert or worker sign-up instead.
Which task type pays best?
Model evaluation and expert reasoning have both the highest ceilings ($250/hour in our tracked data) and the widest range, because the same task type covers everything from basic comparisons to specialist judgement in medicine, law or finance. Transcription and voice or video capture sit lower unless a rare language or a specialist skill is involved. The task label tells you what you’ll be doing; your own credentials mostly decide where you land within its range.